EasyEdit

The de-facto knowledge-editing framework for LLMs (~2.9K★, ACL 2024): ships the KnowEdit benchmark and implementations of most editing methods (ROME, MEMIT, and successors), making it the standard harness for the field. EasyEdit2 (2025) extends the line into test-time behavior steering — safety, sentiment, persona — via activation steering vectors instead of weight edits. From ZJUNLP, led by Ningyu Zhang and Huajun Chen.